climatological evaluation of circulation classifications ... evaluation of circulation...
TRANSCRIPT
Climatological evaluation of circulation classifications
from the COST733 database based on the Kolmogorov-Smirnov test
Radan HUTH, Monika CAHYNOVÁInstitute of Atmospheric Physics,
Prague, Czech [email protected]
What’s new from my last presentation
• v2.0 of the database• all domains• DJF and JJA• both temperature variables (Tmax &
Tmin)• precipitation
Which of my intentions haven’t been fulfilled
• analysis in a gridded dataset• i.e., station data are only analyzed
GOAL• assess the synoptic-climatological
applicability of classifications• i.e., how well they stratify surface weather
(climate) conditions• demonstrate effect of
– selection of the classification method– number of types– sequencing– adding more variables
• 500 hPa height • 500 hPa vorticity• 850/500 hPa thickness
– seasonality of definition
ANALYSIS
• variables– maximum temperature– minimum temperature– precipitation
• 126 stations from ECA&D database• winter (DJF), summer (JJA)• Jan 1961 – Dec 2000
TOOL• 2-sample Kolmogorov-Smirnov test• equality of distributions of the climate
element under one type against under all the other types
-10 -5 0 5 10 15 20 2540
45
50
55
60
-10 -5 0 5 10 15 20 2540
45
50
55
60
-10 -5 0 5 10 15 20 2540
45
50
55
60
-10 -5 0 5 10 15 20 2540
45
50
55
60
-10 -5 0 5 10 15 20 2540
45
50
55
60
-10 -5 0 5 10 15 20 2540
45
50
55
60
-10 -5 0 5 10 15 20 2540
45
50
55
60
-10 -5 0 5 10 15 20 2540
45
50
55
60
x
-10 -5 0 5 10 15 20 2540
45
50
55
60
TOOL
• at each station• types for which the K-S test rejects the
equality of distributions are counted• the larger the count, the better the
stratification, the better the synoptic-climatological applicability
RANKING OF CLASS’S
• at all stations individually: – for each classification: number of rejected K-S
counted– classifications ranked by the %age of rejected
K-S tests (= well separated classes)– higher %age better lower rank
• for each classification: ranks averaged over stations
• area mean rank ranking of the classification
Classifications examined
• 423 class’s in each domain are ranked• only a subset of class’s enter the analysis• omitted are
– subjective class’s & their objectivized versions– original class’s provided by authors (those with ‘o’
in the name)– WLK method– SOM method
• 367 class’s enter the competition
Classifications examined• 4 methods with 3 classifications, differing in
– number of types (9, 18, 27)• 6 methods with 30 classifications, differing in
– sequencing (no x 4 days)– additional variables (Z500, THICK850/500,
VOR500, all together)– number of types (9, 18, 27)
• 5 methods with 35 classifications, differing in – as above– seasonal definition
• infrequent types (frequency < 10 days in the given season) are omitted
Result 1: Ranking of methods
• area mean ranks averaged over 3 realizations with different numbers of types (~9, ~18, ~27) of each of 15 methods
• result: order of the method, independent of the number of types
Result 1: comparison of methods
… there’s no clear winner
• ranking of methods differs– between variables– between domains– between seasons
0 4 8 12 16
GWT JCT LIT
KRZ PXE PCT PTT LND KIR
ERP CKM CAP PXK SAN RAC
0 4 8 12 16
GWT JCT LIT
KRZ PXE PCT PTT LND KIR
ERP CKM CAP PXK SAN RAC
0 4 8 12 16
GWT JCT LIT
KRZ PXE PCT PTT LND KIR
ERP CKM CAP PXK SAN RAC
0 4 8 12 16
GWT JCT LIT
KRZ PXE PCT PTT LND KIR
ERP CKM CAP PXK SAN RAC
0 4 8 12 16
GWT JCT LIT
KRZ PXE PCT PTT LND KIR
ERP CKM CAP PXK SAN RAC
DJF
0 4 8 12 16
GWT JCT LIT
KRZ PXE PCT PTT LND KIR
ERP CKM CAP PXK SAN RAC
JJA
Variables: ranks averaged over domains
0 4 8 12 16
GWT JCT LIT
KRZ PXE PCT PTT LND KIR
ERP CKM CAP PXK SAN RAC
DJF
0 4 8 12 16
GWT JCT LIT
KRZ PXE PCT PTT LND KIR
ERP CKM CAP PXK SAN RAC
JJA
ranks of Tmin, Tmax close to each other precip somewhat different
larger spread of ranks, differences even between Tmin x Tmax pattern more chaotic firm conclusions hard to draw
Domains: ranks averaged over variables
0 4 8 12 16
GWT JCT LIT
KRZ PXE PCT PTT LND KIR
ERP CKM CAP PXK SAN RAC
JJA
0 4 8 12 16
GWT JCT LIT
KRZ PXE PCT PTT LND KIR
ERP CKM CAP PXK SAN RAC
DJF
• no apparent geographical dependence• sensitivity to the size of the domain
• sensitivity to the domain size apparent for more class’s• some regional dependence
Overall rankingmethod DJF JJA
GWT 2 6 JCT 8 12 LIT 3 4 KRZ 7 8 PXE 11 5 PCT 14 13 PTT 10 15 LND 15 11 KIR 9 14 ERP 13 10 CKM 1 1 CAP 5 3 PXK 12 2 SAN 4 9 RAC 6 8
Short summary
• rankings vary among target variables, across domains, between seasons
• caution: results are contaminated (potentially biased) by unequal numbers of really occurring (enough populated) types; the contamination is stronger in JJA
• several well-performing methods can be identified: CKM, CAP, LIT, GWT
• several methods cannot be recommended: PCT, PTT, LND, KIR, ERP
Result 2: effect of sequencing
• all pairs of classifications– differing in sequencing (no vs. 4-days)– with all other attributes equal
• difference in rank is calculated• histogram of differences• t-test: equality of the difference to zero
-240 -160 -80 0 80 160
0
0.04
0.08
0.12
0.16
-240 -160 -80 0 80 160
0
0.04
0.08
0.12
0.16
-240 -160 -80 0 80 160
0
0.04
0.08
0.12
0.16
0.2
-240 -160 -80 0 80 160
0
0.04
0.08
0.12
0.16
-240 -160 -80 0 80 160
0
0.04
0.08
0.12
0.16
-160 -80 0 80 160 240
0
0.04
0.08
0.12
0.16
-240 -160 -80 0 80 160
0
0.04
0.08
0.12
0.16
-240 -160 -80 0 80 160
0
0.04
0.08
0.12
0.16
-240 -160 -80 0 80 160
0
0.04
0.08
0.12
0.16
-240 -160 -80 0 80 160
0
0.04
0.08
0.12
-240 -160 -80 0 80 160
0
0.04
0.08
0.12
-240 -160 -80 0 80 160
0
0.04
0.08
0.12
0.16
EFFECT OF SEQUENCING, DJF
TX
TN
RR
D00 D03 D07 D09
-72 11
-99 11
+14 9
-36 12
-55 11
+122 13
-29 13
-65 13
-12 14
+89 12
-83 13
+63 14
Effect of sequencingDJF JJA
domain TX TN RR TX TN RR00 -72.1 -99.3 +14.1 -75.5 -118.8 +41.401 +4.7 -35.2 +83.0 +40.5 -29.0 +107.102 -59.0 -75.3 +113.0 -13.8 -77.8 +94.703 -35.9 -54.6 +121.7 +2.3 -9.5 +88.804 -6.0 -69.4 +94.1 -1.3 -76.5 +90.505 -21.2 -40.5 +113.0 +6.9 -40.3 +82.906 -33.7 -58.7 +103.3 +30.1 -34.4 +108.007 -29.5 -65.1 +88.9 +2.0 -37.4 +102.008 -34.4 -59.9 +80.9 +9.1 -9.4 +99.909 -11.6 -83.2 +62.7 +22.7 -78.8 +79.310 -29.2 -53.8 +72.2 -8.9 -41.1 +92.411 -14.3 -37.4 +96.9 +42.2 -30.4 +124.7
• improves stratification for temperature• improvement larger for Tmin• deteriorates it for precip• improvement is largest / deterioration smallest for the large domain
• positive effect on temperature is weaker & less ubiquitous• deterioration for precip similar to DJF• same effect for D00 as in DJF
Result 3: sensitivity to the number of types
• all pairs of classifications– differing in no. of types
• 9 vs. 18• 18 vs. 27
– with all other attributes equal• difference in rank is calculated• histogram of differences• t-test: equality of the difference to zero
Result 3: sensitivity to the number of types
• significantly better results for lower nos. of types in almost all cases
• the only few (=3) exceptions, all for JJA and for the comparison of 9 vs. 18 types
Result 4: effect of additional variables
DJF 500 hPa height 1000/500 hPa thickness 500 hPa vorticitydomain TX TN RR TX TN RR TX TN RR
00 -56.0 -58.4 -77.7 -69.5 -73.7 -76.9 +12.3 -14.6 +11.501 -31.7 -39.6 -8.4 -34.5 -45.7 +10.8 -2.5 -22.4 +19.102 -29.7 -52.3 -16.4 -30.5 -47.5 +17.9 +13.7 +11.3 +2.703 -55.0 -60.3 -32.5 -28.3 -28.2 +3.4 +29.2 +34.0 +22.804 +8.3 +23.2 -10.4 -7.7 +.6 -7.7 +25.5 +33.1 -9.405 -63.1 -54.1 +4.6 -13.5 -9.6 +28.8 +29.9 +41.2 +47.706 -67.8 -66.6 -19.9 -29.2 -34.9 -16.8 +10.0 +30.4 +5.807 -15.9 -29.3 -2.3 +3.7 -28.0 +8.1 +34.9 +25.7 +10.508 -77.7 -89.4 -30.1 -66.9 -85.8 -23.0 +32.8 +21.1 +22.909 -72.0 -34.4 -25.0 -38.1 -10.3 -2.1 -23.9 +28.6 +26.010 -76.0 -92.5 -28.1 -39.2 -70.3 -20.5 +27.0 +28.2 +11.511 -75.8 -78.1 -82.2 -62.0 -67.1 -61.0 -6.3 +15.1 +13.2
adding height or thickness:• improvement for temperature –though spatially variable• varied response for precip• height more effective than thickness
adding vorticity:• general deterioration
Result 4: effect of additional variables
JJA 500 hPa height 1000/500 hPa thickness 500 hPa vorticitydomain TX TN RR TX TN RR TX TN RR
00 -60.5 -78.4 -36.6 -42.4 -54.1 +19.6 -5.5 -.7 -4.801 +31.3 -74.9 +71.5 +61.1 -33.9 +87.4 +10.5 -14.9 +47.002 -36.2 -98.2 +86.7 -5.6 -88.9 +92.8 -33.6 -27.4 +1.803 -48.1 -107.9 +51.9 -17.2 -85.8 +52.3 -66.2 -50.6 +20.104 -19.9 -51.8 +56.5 +29.2 -53.3 +107.0 +15.5 +14.2 +4.205 -65.1 -124.6 +67.2 -12.2 -90.9 +79.5 -38.2 -16.9 +12.006 -151.0 -183.6 -16.3 -85.5 -148.7 +28.6 -61.8 -29.9 -1.507 -119.0 -157.7 -12.8 -96.4 -136.3 +12.2 -44.7 -22.7 -18.508 -137.1 -141.4 -16.5 -109.2 -136.9 +14.7 -37.0 -7.5 -7.509 -83.3 -126.1 -123.6 -80.3 -132.5 -87.3 -8.3 -.8 -42.610 -162.6 -172.5 -27.5 -148.2 -172.5 -27.5 -42.0 -14.7 -11.711 -104.6 -92.2 -61.5 -99.4 -101.4 -35.0 -21.5 -5.7 -17.7
adding height or thickness:• improvement for temperature; stronger in SE half of Europe; stronger for Tmin• deterioration / improvement for precip in N+NW / S+SE Europe
adding vorticity:• improvement for temperature in most domains• varied response for precip
Result 5: Effect of seasonality
DJF JJAdomain TX TN RR TX TN RR
00 -44.0 -71.2 -46.4 -132.7 -94.5 -156.401 -40.6 -28.2 -17.1 -44.2 -54.7 -92.602 -75.2 -63.5 -35.1 -46.5 -14.6 -87.003 -15.5 -25.7 -55.4 +3.4 -17.0 -58.104 -37.8 -45.3 -36.2 -62.2 -49.4 -20.105 -42.5 -32.6 -21.9 -14.9 -29.0 -58.806 -26.1 -35.8 -30.5 -12.3 +18.6 -9.407 -27.1 -39.4 -28.9 -52.5 -39.5 -55.008 -36.8 -47.5 -19.6 -44.7 -57.3 -40.109 -38.1 -56.4 -61.6 -90.7 -84.0 -15.710 -16.9 -25.7 -25.5 -70.7 -64.5 -45.111 -44.9 -45.5 -17.0 -18.8 +36.9 +8.2
• general improvement• some geographical variability
BUT:systematic difference in the no. of types (7 seasonal vs. 9 non-seasonal) improvement may be partially an artifact of this difference
instead of CONCLUSIONS
• what else might have been done• or has been done by someone else and
might be nice to be combined with this study
• gridded dataset (Ensembles, NCEP or ERA40)
• other criteria of stratification